PH热榜 | 2026-05-05

Looking forward to seeing what you're building with @Kilo Code!

Hey Product Hunt 👋 Brian from Kilo Code here.

We just shipped the biggest update to our VS Code extension since launch. The entire thing has been rebuilt on OpenCode server, which is the same open-source core that powers our CLI and Cloud Agents. One engine across every surface, so when we improve something, it gets better everywhere.

The headline feature is real parallelism. Kilo can now run multiple tool calls at the same time (file reads, searches, terminal commands all firing concurrently), and it can spin up parallel subagents that each handle a piece of a larger task simultaneously. You actually feel the speed difference.

A few other things shipping in this release:

• Agent Manager — run multiple independent agents in separate tabs, give each one a role, and use git worktrees so they never step on each other's code

• Inline code review — leave line-level comments directly on agent diffs, just like a PR review, and send them back as structured context

• Multi-model comparisons — run the same prompt through different models side by side and pick the best result

• Cross-platform sessions — start in the CLI, pick up in VS Code, share with a teammate

Kilo is open source, runs 500+ models at provider cost (zero markup), and has over a million developers using it. We'd love for you to try it out and tell us what you think!

Came from Roo Code a few months ago and honestly haven't looked back. The migration was smoother than expected. Good luck for today!

Hey Product Hunt 👋 - Job from the Kilo team.

Very excited for this launch. The new Kilo for VS Code is my daily driver, and this rebuilt version with agent manager is in my opinion the next step in agentic coding. I now let multiple agents run at the same time using agent manager, and it speeds up my workflow a lot. Super curious to hear what you think of it!

🚨 We're also hosting 2 live sessions TODAY:

I’ve been using Kilo Code for a while, and this update feels noticeably faster.

The parallel tool calls are the part I felt right away. It doesn’t sit around waiting as much, especially on bigger tasks where it needs to search files, read code, or run commands. The Agent Manager is also really nice if you’re juggling a few things at once without wanting everything mixed together.

Inline review on diffs is probably my favorite addition. It makes giving feedback to the agent feel a lot closer to how I’d review a teammate’s PR.

Overall, this is a really strong update. Kilo is becoming one of the few coding tools I actually keep coming back to.

Been using v7 for a while, and even the versions before this. Took a while to get used to, but generally excited to see it being built on OpenCode. Can't wait for more features especially /remote 🤞🙏

The parallel subagents with git worktree isolation is the part that actually makes sense to me. Every other tool just runs agents on the same files and hopes for the best. Congrats on shipping this!

If one agent is refactoring an API and another agent is consuming that API, how do they handle the dependancy? Do they share context live or waiting for human reviews?

How does the multi agent system comparison handle tokens ? does it run them all in the background simultaneously ? btw Congrats on the launch :)

Used the old Kilo VS extension and have been using the new. Love the changes and it works smoothly. Excited to see many of the coming updates, too.

Been loving my experience with Kilo! Love the team and the speed at which things are shipped. This is just another great example of the speed at which this team can produce great work!

Been on Cursor for a while but the model lock-in is starting to bother me. The 500+ models angle here is hard to ignore. Does switching models mid-project break any context?

Using Kilo for Resume Matcher. Works really well, however, ƒor this launch. I'd suggest some features that Kilo should add. The first one is Skills, just like Claude-Code. A .kiloignore, and custom routines to be fired up to check for dependencies, security risks, and other supply chain attacks that may be in transitive dependencies. Because the more we vibe-code, the better the security should be, and the more exhaustive the reviews should be.

Niiiice! Looks amazing and clean

Does the agent manager have any limits on how many parallel agents you can run at once, or is it just constrained by your machine's resources?

WooW this feels really fast.

Great update guys! Nice work!

Congratulations!

Rebuilt from scratch on OpenCode server and still GA'd,  that's not a small thing. Most teams would've shipped a half-baked beta and called it done.

The line level review comments on agent diffs is a really smart UX call. Feels like the missing link between AI wrote this and I actually trust this going to prod.

How does Kilo handle context limits when you've got multiple subagents running on a large codebase? Does each subagent get its own context window or do they share?

Just curious, when multiple subagents are running in parallel, how does the merge back to the parent agent work? Does it ever create conflicts when two agents touch overlapping parts of the codebase?

been looking for something like this, the ai coding space is getting crowded but this looks focused

Hey Product Hunt community! 👋

I'm Ajay, co-founder & CTO of Velo. A few weeks ago, we launched Velo 1.0 here - an AI tool that turns a raw screen recording into a polished, share-ready video message.

The response blew us away. But one thing kept coming up in the feedback: "It works, but it doesn't feel effortless yet."

That stayed with us. For a tool built to make communication easier, "not effortless" is the whole problem.

So we went ideated, iterated, and are finally shipping Velo 2.0 here today.

Here's everything about Velo 2.0

• Chat-native interface: Chat with Velo to shape it into a polished video - add narration, edit scripts, add voice clones, effects, and more, all through conversation.

• Streaming processing: Processes as you record, noticeably faster previews

• Create video messages + documentation: Record once, get a polished video and a structured written document out of the same session

• New voice model: Emotive, human, captures tone and energy, not just acoustics

• Silent video handling: No audio in your video? Velo understands the context and generates the script instead of failing

• Video to Doc: One click turns any video into a written article

• Context-aware script rewriter: Set your persona and audience, Velo rewrites the narration

We've been using 2.0 internally for weeks. It's the first version where I make a video and don't feel like something's slightly off.

Please try it out and let us know what your thoughts are: usevelo.ai

We're all ears. Thank you for your support.

Hey PH community

The mission behind Velo has always been the same - help people communicate better through video, without needing to be a video person.

No camera setup. No re-recordings. No editor. Just you explaining something clearly, and AI doing the rest.

Velo 2.0 gets us meaningfully closer to that. The chat-native interface means anyone can start without a learning curve. The new PDF flow actually captures what you want to say, not just what's on the slide.

The voice model sounds like a human having a conversation. And features like Take Control and Video to Doc open up use cases we couldn't touch before.

What can you expect when you try it:

• A starting experience that guides you through Velo

• Videos that sound like you - not a robot reading your script

• PDFs that become narrated videos in minutes

• A written doc from every video, with one click

• Faster previews, sharper output, less waiting

We're a small team building something we genuinely believe in. Every launch on PH has shaped what Velo is. This one is no different.

Try it. Break it. Tell us what you think. We're here all day and open to all your feedback and thoughts.

Appreciate your support.

And to everyone who used V1 and stuck around - thank you.

Congrats on the Velo 2.0 launch. Really liked what the team is building here.

I’ve personally struggled with making videos because the actual recording is only one part of the work. The harder part is everything after that - trimming, fixing awkward pauses, making the flow better, adding narration, creating a doc, and then making it look shareable.

That’s where Velo has been genuinely helpful for me. It makes video creation feel much less like editing and much more like just communicating clearly.

it’s been about a month since our first launch. this ended up being more of a rebuild than we expected. we reworked a lot of core parts to make things faster and flow better, and pushed it more towards an AI native way of using it. curious what you all think.

Velo 2.0 is heavily shaped by user feedback and analytics from 1.0, especially around where things felt slow or not “worth sending.” A lot of the changes come directly from that. Very excited to see how it feels for everyone and would love honest thoughts.

This version is built directly on top of user feedback. After launching 1.0, one theme kept coming up: it works, but it doesn’t feel effortless yet. That became the entire focus for V2.

We went deep on listening, analyzing where the friction actually was, and improving the core experience not just adding features. All of it ties back to our mission: helping people communicate clearly through video, without needing to be “video people.”

You’ll see that reflected across the product—from more natural and expressive voice clones, to context-aware script writing that adapts to your audience, to an overall workflow that feels faster and more intuitive end-to-end.

It's been just about a month since we launched Velo 1.0, since then, we've been reading all your comments and feedbacks, last one month, it was a crazy ride. Sleepless nights, complete revamps, all flows getting changes, we gave it all. The first product was usable, but we wanted it to be the fastest and easiest product for all our end user's. In the AI-native generation, our product felt a bit old, so here we are just after a month, the complete product being revamped to an AI-native product. Your feedbacks have been gold for us and we tried to forge the product according to your feedbacks, do give it a try. Hope you all will love it too!

Here again to support for round 2. Cheers, congratulations on shipping improvements so quickly, speaks volumes about the team at Velo.

Been recording walkthroughs for the team and retyping the same stuff into notion forever. Does it handle longer demos cleanly or is it tuned more for short messages?

what happens when the pdf you're turning into a video has a lot of charts and diagrams? does it describe them, skip them, or try to animate them somehow?

congrats!!

Congrats on the launch! Just generated a video! Is it possible to edit the video after it's done?

@ajaykumar1018This is really cool, congratulations! Love that it sounds like the main focus is making it feel effortless. Making videos is not necessarily my forte and honestly takes me forever so I am looking forward to giving it a try.

How does Velo treat the screen recording? Does it do any sort of animation or retouching, or just pure screen capture?

Chat-native editing on raw screen recordings is a meaningfully better mental model than timeline editing for B2B creators. I produce a finance podcast on the side (the ModeLoop podcast) and the bottleneck is never the conversation — it's the post: trim, intro, captions, clip extraction. Voice cloning + script rewriting could collapse 80% of that into a single chat turn for podcasters who want to repurpose long-form audio into 60s teaser videos. Curious whether Velo handles audio-only sources or strictly screen + voice, and whether the script editor preserves speaker turns?

Congrats on 2.0, team!! The frist time I tried V1 I liked the idea but it felt a bit slow. Really glad to hear the streaming processing was a focus this time around. Going to test this today.

the chat-native editing is what got me. curious how it handles more complex edits - like if i want to cut a 3-minute section and re-order two others, is that still a conversation or does it get clunky?

congrats!

@ajaykumar1018 Congratulations. And happy product launch new verson.

I tried Velo and it made creating video walkthroughs really fast and simple. The AI editing and voice features save a lot of time, especially when you don’t want to re-record again and again.

👋 Matt here, co-founder of Flowstep.

Last year, we shipped Flowstep as an AI designer. But you kept telling us the same thing:

"it gets me 80% there, then I rebuild the whole thing in code"

We fixed that.

Flowstep is now an AI design engineer.

→ Prompt or manually edit on an infinite canvas

→ Get production React + TypeScript + Tailwind + shadcn

→ Export the code, or pipe it into Cursor / Claude Code / Windsurf via MCP

Why this exists: The process of fast iteration, multiple concepts next to each other, full control in an infinite canvas 🤌 You shouldn't have to give up that experience because of a translation layer. That's why design and code are the same thing in Flowstep.

Free to try at flowstep.ai, paid plans available from $15 p/m.

💬 Drop a comment — tell us what's missing. We listen and ship multiple times per day.

to answer your question: the one thing usually missing in these tools is state management. if flowstep can help scaffold the basic useState or useEffect logic while i'm building the ui on the canvas, i’m never opening figma for web work again. definitely checking out the free trial today.. @clannachan

Great looking UI, will test it out 🔥

wooooow so cool !!

Wishing GL with the launch :)

Congrats on the launch, good luck team! 💪🏻

Curious how clean the exported production code in real world scenarios .If it's truly usable without heavy cleanup this could be a game changer.

Love the canvas first approach .Most tools still feel constrained but this seems much closer to how designers developers actually think and iterate.to

This is the kind of shift that actually matters less about designing screen more about desigining shippable system.Curious how reliable the generated code is in real projects.

.

The MCP integration angle is interesting feels like you're not just building a design tool but plugging into a larger AI workflow ecosystem.

This feels like a real step toward closing the gap between design and development . The what you design is what ships” promise is exactly what many teams have been waiting for.

sounds like something that actually helps agents to build an authentic UI rather than just vibecoding it!

Yeah! Building fast sometimes means it's harder than ever to get a consistent and cool design. Even more if you need to update core design parameters and I'm sure Flowstep gonna help a lot on it

the '80% there then i rebuild in code' line is exactly why i stopped opening figma. running claude code daily and the design to prod handoff has been the biggest tax on shipping fast as a solo founder. mcp export into the coding agent is the right shape. trying it today !

Thanks flowstep team for making this.
It really helps me to generate different design directions in very short amount of time and 2x my past workflow

Is this meant for legit designers and devs, or teams that are trying to cut corners by not hiring design and dev? Are there guardrails to keep the designs "good".

Can you import a design system?

Hey PH 👋

Alex here, founder of Waydev. Thanks @rajiv_ayyangar for the hunt — and shoutout to the team at Deel for including us in Pitch by Deel NYC. Excited to be part of it.

Nine years ago I started Waydev because engineering leaders were flying blind on what their teams actually shipped. Now half the code is written by Cursor, Claude Code, Copilot, and Devin — and the question from CFOs is the same one in different clothes: is the seven-figure AI bill actually paying off?

Most teams don't know. They're guessing.

Waydev is the measurement layer for AI-written code. We track three things across the full SDLC:

• AI Adoption — who's using which tool, how often, how deeply

• AI Impact — does the AI-written code ship, get reverted, or rot in PR

• AI ROI — dollars in vs. throughput out, across humans, copilots, and autonomous agents

Today we're shipping three things that change how engineering leaders interact with their data:

• 🤖 Ask Waydev — our agent. Ask in plain English, get a real answer. No more dashboards no one opens.

• 📄 Skills — configure Ask Waydev with SKILL.md files.

• 🔌 Waydev MCP — your engineering feed, exposed to any external agent. Plug it into Claude, Cursor, your internal tools — whatever's already in your stack.

The framing we keep coming back to: MCP is the data out. Skills are the instructions in. Ask Waydev is the conversation in the middle.

The bet: engineering leaders would rather have conversations than dashboards.

We've been covered by TechCrunch, TNW, and DevOps.com along the way.

I'll be in the comments all day. Roast the product, ask the hard questions. That's how this gets better.

— Alex

This is super relevant right now. Everyone is pouring money into AI, but barely anyone knows what it’s actually returning. Love the shift from dashboards to conversations, feels way more natural for how teams want to work.

the 'ai adoption vs ai impact vs ai roi' split is the right decomposition. most teams collapse all three into 'are devs using cursor' which tells u nothing about whether the code actually shipped or got reverted. measuring the impact layer is where this gets interesting

The "AI Impact" piece is the hardest problem in this space and most tools quietly fudge it. Devs don't write code in clean buckets. They use Cursor to draft, rewrite half by hand, accept Claude's completion on three lines out of fifty, paste a ChatGPT snippet and edit it for an hour. What's Waydev's actual attribution method when AI authorship is partial, mixed, and deniable? IDE-plugin self-reporting, statistical fingerprinting on commits, or something else?

Hi everyone! Renee here - one of the builders of Ghostwriter 👋

Really excited (and a little nervous) to share this with you all. A bit of context on the story behind Ghostwriter:

1.5 years ago, I quit my chief of staff job to invest in myself and learn how to build. Around the same time, I started posting on LinkedIn and X for the first time - just sharing what I was going through and learning.

It ended up completely changing my trajectory! I started getting messages from people I looked up to, and even job offers in my DMs. I realized how powerful it is to just show up and post content online.

But… staying consistent is hard (and it's time consuming!).

So Nathan and I built Ghostwriter for ourselves - something that helps you create content that actually sounds like you, without it feeling like a chore.

Personal branding is probably the most high leverage thing you can do for yourself right now. Ghostwriter helps you create a personal content engine that works for itself.

With Ghostwriter, you can:

∙ draft a month's worth of posts that sound like you

∙ customize your tone of voice

∙ schedule & publish directly to Linkedin & X

No more asking ChatGPT to “make it punchier", now you get your content handled end to end. Even my dad uses it 😇

Our early testing has been loved by hundreds of users, and we're excited to finally open it up to you all!

Hope you enjoy it, I'd love to hear what you think!

Try for free → https://ghostwriterrr.com/

Thank you!

Renee

⊹⊹⊹⊹ a few more links ⊹⊹⊹⊹

→ Join our community (we're friendly I promise): https://tinyurl.com/2eyt62nj

→ Follow us: LinkedIn & X

→ My socials: LinkedIn & X

Curious how well it adapts to different tones over time .Does it learn from past posts or require manual tweaking each time?

Interesting positioning as an end system instead of just a writing assistant . Drafting scheduling publishing in one flow makes a lot of sense.

the 'consistency is the hard part' framing is the real insight here. tools have never been the bottleneck for posting, taste and reps are. how does ghostwriter handle the cold start problem when someone has barely posted before and there's no real voice to mimic yet?

Looks super cool, can't wait to try it out. Would LinkedIn have a way to detect which posts are human written and which ones are not though? And would that impact the algorithm?

Seems really useful! Sometimes I am just too tired to come up with ideas on what to write. Also, are you planning to add support for Threads?

Awesome stuff! I know a few friends who case use this :)

so keen to give this a whirl, as someone who posts a lot on different platforms, this will be a gamechanger.

ps. such a cute logo!

This is really sick! 🔥

Been using Ghostwriter for my Linkedin posts last few months as an early beta user - has been game changing!

Really good stuff @reneezhang @nathanksou

Hey folks! 👻

Thank you for all the support! We're so excited to bring Ghostwriter to you.

Personal branding is the most high-leverage thing you can do for yourself right now, especially as a builder/founder.

Ghostwriter helps you create a personal content engine that sounds like you, without feeling tedious.

I started building my personal brand with Ghostwriter's help while building out the agent and platform. It's been incredible to watch it grow. I'm excited to see how your Ghostwriter writes for you.

Try for free: https://ghostwriterrr.com/

Join our community: https://tinyurl.com/2eyt62nj

Nathan

Huge thanks to @rajiv_ayyangar for hunting us 🙏

Hey Product Hunt 👋

Julien here, co-founder of Oriane (with Yuri and Thibaut).

Quick story. Two years ago, Yuri was running product at Jellysmack working with creators like MrBeast. They kept asking him: "What's my actual presence on TikTok? Where are my clips being remixed? The data simply didn't exist.

I was leading growth at DRESSX with clients like Lacoste, Adidas, Tommy Hilfiger... Same problem. Still the same at LVMH today, where brand teams rely on social listening tools like Brandwatch or Meltwater, that treat video like text. Captions and metadata aren't the content anymore.

So we built the perception layer. Oriane watches millions of TikToks, Reels, and Shorts a day, and turns what's on screen, in audio, and in captions into structured intelligence. We're the eyes and ears of the AI stack.

Brain → LLMs.

Eyes & ears → Oriane.

A few things about us:

🍒 Thibaut, our CTO, took our cost to vectorize 100K videos from $60K to $60. That's a 1,000x reduction. It's why this is even a viable category.
🍒 Customers include McCann, LVMH, Woo, Dior, Disney, Pierre Fabre, Estée Lauder, and more.

What I'd genuinely love feedback on:

🍒 If you work at a brand or agency, what's your current workflow for tracking video mentions, and what does it cost you?

🍒 What would make Oriane essential for your team, not just useful?

🍒 Is "the eyes and ears of the AI stack" the framing that lands for you, or is something else clearer?

Special offer for the PH community: 1 month free on Pro plan ($499 value) with code HUNTER-499. Good through May 30.

Yuri, Thibaut and I are here all day.

Bring your toughest questions 🙏

Yuri here, co-founder/CPO 👋

The Jellysmack story Julien mentioned, that's the one that made me sure this had to exist.

MrBeast asking me where his audio was being remixed, and the honest answer being "we genuinely cannot tell you."

I never want to give that answer again. If you're a creator or work with creators, I'd love to hear what your version of that question is.

We probably haven't built the answer yet, but we will.

Thibaut here, CTO 👋

Quick note on the cost reduction Julien mentioned, happy to nerd out on the infrastructure side if anyone's curious. The short version: the whole industry approaches video understanding as "extract everything, then index",  we flipped it. We took the cost to vectorize 100K videos from $60K to $60. That's 1,000x. It's why this category is suddenly viable.

Engineers in the comments, fire away.

Traditional scrapers and social listening tools hunt for hashtags or captions, but they are deaf and blind to what actually happens inside the frame. This platform watches and transcribes social video content, surfacing brand mentions and visual trends that never make it into a text-based description.

It uncovers the specific micro-influencers who are already organically discussing a niche, rather than just the high-priced creators with the right keywords in their bios.

The real utility lies in the data structure. You can analyze video content to deconstruct visual hooks and audio patterns and determine exactly why specific segments engage viewers. Feeding this granular output into an LLM allows for the creation of trend reports that feel like they came from a high-priced agency, but in a fraction of the time.

While the speed is impressive, users should anticipate a learning curve. Navigating the sheer volume of "visual mentions" requires a disciplined approach to filtering, or you risk drowning in data. It isn't a "set it and forget it" solution; it demands an active strategist to turn these observations into a coherent content plan. That said, the team behind Oriane is ready to assist, and the website includes tons of case studies and use cases to inspire you. In summary, it offers a depth of category intelligence that other tools simply cannot match.

Very interesting. On the technical side, how do you manage to have such low latency with so many videos and what I imagine is a vector search? You used pinecone or qdrant?

"Captions and metadata aren't the content" should be on a t-shirt.

That sentence is the entire wedge against legacy social listening. Every PMM working in brand intelligence is going to steal it within the month. Watch.

This sounds interesting for creator research. Which platforms are you pulling videos from right now, like TikTok and Reels or more YouTube Shorts?

The watches millions of video a day claim is wild .Would love to understand the infrastructure behind that.

Comparing this to tools like Brandwatch and Meltwater is smart they really do treat video like text.

The Jellysmack creators ecosystem origin story makes this much more credible. You’ve clearly felt the pain firsthand.

Habéis considerado un plan de pago por videos procesados?

Hi Yuri / Thibaut, what other dimensions on top of AI Vision and audio transcripts do you plan to release then? Music ?

Just gave it a spin on the free version, looks pretty solid. How do you manage to vectorize so much videos without running out of money?

the 1000x cost reduction on vectorization is the actual unlock here. built a tiktok audio trend pipeline last year for a content channel and the math on doing it at scale across millions of videos was brutal at the time. brandwatch treating video like text has been the wedge sitting there for years, surprised it took this long for someone to go after it properly

@julien_rosilio Congrats on your launch 🙌

This feels like a natural evolution. Social listening went text → sentiment → image recognition → video intelligence. Each step was inevitable.

Curious how deep the analysis goes are you just tagging content or actually understanding hooks, pacing, and storytelling patterns?

congrats on the launch, oriane team! how did you cracked inference problem?

Truly love the product! As I’m getting started with IG, Oriane helped me to figure out what content to build and what works! Thanks for building this @julien_rosilio and @yuri_mihaileanu1 🙌🙌🙌

@julien_rosilio Watching this from the music side. The luxury wedge makes sense

as a beachhead, but the perception layer applies to everything.

We have the same "captions don't capture the content" problem,

just with audio fingerprinting, sample tracking, and unauthorized

derivative use.

Question for the team: if you were going to expand the perception

layer beyond brand listening, which adjacency would you pick

first? Music IP detection? Sports broadcast monitoring? Live

event surveillance? Curious how you're thinking about the second

beachhead.

Hey Product Hunt 👋! I’m Faisal, one of Intuned’s co-founders.

Today we’re launching Intuned Agent: an AI agent that builds and maintains production browser automations.

What makes Intuned Agent different is that it is built into Intuned’s browser automation infrastructure. It can use platform capabilities like auth, stealth, CAPTCHAs, proxies, schedules, retries, and observability, then inspect failed runs, traces, logs, and screenshots to debug issues and write fixes when websites change.

When we first launched Intuned, it was a code-first platform for building and running browser automations.
But almost immediately, customers started asking us for the same thing:

“Can you just build and maintain these automations for us?”

So we did. Over the past year, we ran a services motion alongside the product and processed 40M automation runs (20M mins) for a limited set of customers.

We spent more than a year trying to turn that “solutions engineer” workflow into product. The unlock was embedding Claude Code through the Claude Agent SDK directly into Intuned.

The most exciting feature that Intuned Agent unlocked is self-healing. When a project fails, if self-healing is enabled, Intuned can detect the issue, the agent can inspect the failed run, diagnose what changed, write a fix, and redeploy! with you controlling how much autonomy it gets.

That’s the core idea behind Intuned Agent: you still get real code you can inspect, own, and run in production, but the agent helps with the painful parts of building, debugging, and maintaining it.

To make this more concrete, we made a 5-minute video walking through Intuned, Intuned Agent, and how the platform works: watch it here.

Would love feedback from anyone building browser automations in production. Head over to Intuned and start building for free!

Hey Product Hunt!

Nasser here, part of the team behind Intuned Agent.

Building Intuned Agent took us way longer than we expected to get right.

Our first attempt was a setup where the chatbot would collect requirements, then hand everything off to a rigid pipeline. One step to discover the site, another to structure the data, and a final step trying to group everything and fix whatever broke along the way. It looked clean on paper, but fell apart on real websites.

The biggest signal was that our own solution engineering team (who help customers build and maintain automations) didn't want to use it. It was too rigid for the kind of messy problems they deal with.

We tried to improve that setup for a while and learned a lot, but it became clear the approach itself was the problem.

When we started using Claude Code for generic coding tasks, we liked how it worked. Its fluid, end-to-end approach felt like the right direction for us to adopt. That's what pushed us toward rebuilding with the Anthropic Agent SDK.

We took all the browser automation knowledge we'd built over time and turned it into reusable skills, then let a single agent drive the process instead of forcing it through a fixed path.

Hey Product Hunt! I’m Rauf, one of the engineers behind Intuned’s agent runtime and UX.

I’ve spent the last 3 months focused on building the user experience for Intuned Agent. One thing became obvious quickly: the hard part is not getting the model to take actions, it is making the entire workflow understandable, controllable, and reliable.
An agent like this is never just “a chat.” There is conversation state, code state, browser state, session state, billing state, and human approval state, all moving at once. The UI has to make that legible: dense enough to be useful for builders, calm enough that it does not become stressful.
Most of the real work is the product layer around the model: session runtime, queueing and turn control, reconnect and resume flows, interruption handling, human-in-the-loop approval, sandboxing, and recovery paths for when the agent stops, stalls, or crashes mid-task.
The standard I kept coming back to was simple: can people actually use this with confidence, and can we sleep while they do?
The visible surface is “an AI agent that helps you build browser automation.” The real work was making it dependable enough that builders would use it on a real project, not just admire it in a demo.

Hey Product Hunt! I’m Omar Bishtawi, and I lead the platform team here at Intuned.

One thing that makes Intuned Agent different from a normal chat-based coding agent is the kind of work it has to do.

Agent sessions can be long-running. They drive real browsers, inspect live websites, write code, run tests, debug failures, and sometimes keep working after the user has walked away. That creates a very different infrastructure problem: sessions need to be resilient, interactive, fast to start, and secure.

A few things we built to support that:
Resilient sessions
 Agent sessions are dynamic and unpredictable in their resource needs. We continuously checkpoint session state so the agent can resume mid-task after any interruptions like machine failures, OOMs, network hiccups, or client disconnects.
Live feedback that survives reconnects
 We stream the live browser and agent activity in real time, while making the session reconnect-safe. Close your laptop, open it again, and you can pick up where you left off.
Fast startup without wasting compute
 Idle machines cost money, but cold starts hurt the experience. We use microVMs and adaptive warm pools to get from “click run” to “agent working” in a couple of seconds, without keeping too much idle capacity around.
Security end to end
 We think a lot about what it means to let an agent act on your behalf. That includes how we store shared information, limiting the agent’s access, sandboxing actions, and running each agent session in its own microVM with a dedicated kernel.

Most of this infrastructure is invisible when it works, and very visible when it does not. Excited to see what people build with Intuned Agent.

Hello everyone, I’m Omar, a software engineer at Intuned.

When developers build browser automations in Intuned, they do more than write code. They test against live sites, debug failures, deploy updates, monitor runs and jobs, and inspect browser traces and logs. We wanted Intuned Agent to be able to do that same workflow inside Intuned.

We already had pieces of this across the dashboard and APIs, but they were built for humans and apps, not for an agent trying to complete an end-to-end workflow. The agent needed one interface that could manage the platform, discover what actions were available, and operate reliably across build, debug, deploy, and maintenance tasks.

That led us to build the Intuned CLI.

With the CLI, the agent can operate Intuned through the same tool engineers use. It can explore commands with --help, discover the right workflow, run jobs, inspect results, and interact with the platform in a predictable way.

We also added a hook system so we can tailor behavior for the agent, provide guidance, and handle specific scenarios without making the core interface messy.

The result is a cleaner interface for the agent, and a more powerful workflow for engineers building with Intuned.

Get your frist scraper, RPA or crawler up and running within minutes with the Intuned Agent!

Hey there Product Hunt 👋

I'm Izzat, one of the engineers building Intuned Agent.

Anyone who has worked on browser automation knows how quirky things get: iframes, CAPTCHAs, selectors that break the next day, infinite scroll, logged-in sessions that expire...

What makes Intuned Agent exceptional is the browser-native harness we've spent the last few months building around it. The agent handles all of these cases naturally while writing the automation code.

Customers love it, and it keeps surprising them. It happens regularly where the agent spots a hidden network endpoint mid-exploration and turns what would've been a few-hour scraping task into a single API call. This is one example of many.

We're offering free credits to try it out. Give it your most challenging browser task and let us know what happens 🙌

I’m Ahmad, co-founder of Intuned, and I lead the team working on Intuned Agent.

Customers use Intuned Agent to build all kinds of browser automations, all running on Intuned’s production infrastructure with proxies, stealth, CAPTCHA handling, auth, scheduling, and observability built in.

Here are a few real-world use cases we’ve seen on Intuned:
• Government data extraction across 50 sources for a B2B SaaS data company, including RFPs, solicitations, and meeting minutes. Live in days.
• E-commerce catalog migration for teams launching dozens of stores a week, moving products, prices, images, and descriptions into a clean target schema.
• A tech jobs aggregator that pulls thousands of open roles from hundreds of company hiring pages.
• Insurance quote collection from auto and home insurance providers for a price comparison platform.

This is where Intuned really shines: scale. One customer now runs close to 1,000 automation projects every day. Intuned Agent monitors failures, patches code, and redeploys fixes, so their engineers only review what it cannot resolve.

Drop a public site you’ve been meaning to scrape or automate, and I’ll run it through Intuned Agent. Happy to share what comes out. 

Hey hunters 👋 I'm Mohamed Khalil, Solutions Engineer at Intuned — but before this, I spent ~3 years in the scraping trenches, running pipelines across 5,000+ sites (e-comm, gov procurement, real estate, the usual chaos).

Most of those years were rebuilding the same things over and over: proxy rotation, fingerprint spoofing, captcha flows, the selector that breaks every Tuesday. So when I joined and saw what the agent does, my reaction was very much *"oh, this is the layer I wish I had two years ago."*

It isn't that it clicks buttons — plenty of things click buttons. It's that it handles the boring fight: bot detection, long-running sessions, and recovery when a flow drifts. The stuff that usually eats 80% of a scraping engineer's week.

Excited to see what you all build with it 🚀

Faisal congrats on the launch. self-healing is the thing I keep getting tripped up on in our own browser-automation pipeline. curious how Intuned tunes the trigger threshold. is the detection mostly on assertion failures (selector missing / element not in expected state), or is there a separate heuristic that watches for layout drift across runs even when the script technically completes? false positives waste agent runs, false negatives strand users, and we've found the threshold is the brittle part of the whole loop.

a most 'AI web scrapers' fails because they ignore the boring stuff like proxie, stealth, and retries. building the agent directlly into your existing automaton infra is a smart move. it means the agent isn't just writing code, it's actually managing the execution environment. definitely checking out the claude sdk integration @ahmad_ilaiwi

I've spent way too many hours of my life manually scrolling TikTok looking for products. The "captions and metadata aren't the content" line in the maker comment hits hard.

Question for Julien: how does Oriane handle when a creator shows the product but doesn't say the brand name?

Hey Product Hunt! I'm Charbel, and I built PaceBar because I kept noticing, too late, that I'd been working way too intensely.

A typical session: one task, then Slack, a doc, a few browser tabs, a couple of AI chats, another tab, and suddenly the "focused block" was five things at once. The annoying part was always realizing it after I was already drained.

I didn't want a productivity dashboard or another habit system. I just wanted a small signal in the menu bar that said: your pace is climbing.

So that's PaceBar.

PaceBar is a quiet menu-bar instrument for Mac. It turns local interaction patterns, like activity timing, idle time, focus time, and app changes, into a 0–100 session-load readout that adapts to your working baseline, so you can catch a rising pace before the session gets heavy.

When things are balanced, it stays out of the way. When the session gets heavy, it can nudge you to step away briefly and return to one task. When there's headroom, it can prompt you to protect the next block for the work that matters.

Privacy was a hard constraint from day one: no account, no telemetry, no cloud processing, no privileged access. PaceBar doesn't read your screen, see what you type, or touch your files. Everything stays on your Mac.

It's built for people who spend the day context-switching across code, docs, chats, browsers, design tools, research, and AI tools, and want to notice the pace before it gets away from them. Not to optimize every minute. Just to catch yourself before you overdo it.

I'd love feedback on the core idea: does a quiet menu-bar instrument for session load feel useful in a real workday? I'm especially curious whether the Calm / Steady / High framing reads clearly at a glance, and how subtle the nudges should be.

Very interesting.
Does it give a bit of a data read of last week, month etc?

This looks super interesting.

I wanted to try it out but it's not available in my region, any plans to release it globally? 🙃

Hi Product Hunt.
Shubham here from Firstwork.

3 weeks. That's how long many frontline hiring teams still take to move workers from application to first shift.
Not because recruiting failed. Because operational workflows slow everything down.
Verification. Training. Scheduling. Compliance. Documentation.
Meanwhile:

• Workers drop off

• Recruiters chase paperwork

• Ops teams rebuild schedules

• Compliance teams prepare for audits manually

We built Firstwork to operationalize the entire workflow.

Firstwork automates document validation, verification, onboarding, training, compliance, and scheduling in one clean, operational pipeline.

Workers reach Day-1 faster. Ops teams reduce manual follow-up. Compliance workflows stay audit-ready.

The thing we didn't expect: customers talked about audit readiness almost as much as hiring speed.
One customer told us, "We passed an audit without scrambling for documentation for the first time in years."

If you run frontline hiring or workforce operations, I'd love to hear where your biggest bottlenecks are today.

First time this clicked for me: I watched an ops manager pull up her browser.

14 tabs open, switching between them while talking to a candidate. ATS, e-sign, two state licensing portals, training tracker, scheduler, spreadsheet.

She said "I do this 30 times a day." We collapsed her workflow into one tab. She sighed with relief. That was my aha moment.

The compliance rules engine is the part of Firstwork we're proudest of. The average recruiter was losing about four hours a week to verification checks and manual compliance lookups, and we built the engine to give that time back.

We worked with Firstwork at my last workplace and they cut down the time it takes to review documents and perform browser-based tasks from hours to seconds for us, which had a huge positive effect on cutting down the overall time to onboard new employees. Highly recommend them!

A much needed product in the market! This feels less like another generic AI hiring tool and more like operations infrastructure for frontline teams. I'm excited for the robust integrations, could be a game changer.

Reading licenses with AI was the easier part of V1. The hard part was figuring out that "Sarah Chen" with a Texas nursing license, "S. Chen" with an expired California one, and "Sarah J. Chen" in New York are all the same person. The matching logic took longer to build than the AI did.

We started with solving hiring and onboarding.

Ran into everything around it - scheduling, follow-ups, training, compliance - spread across tools and manual work.
So we built across that layer - automations, calendars, hiring flows, LMS, AI agents, even callers - to make it hold together.

Feels right in parts. Curious where it doesn’t.

Looks exciting. How is this different from Greenhouse, Workday, or Fountain?

the compliance looks promising. But I missed the price to use this for startups

Do you help in hiring international resources? Onboarding international resources is still a huge challenge for us.

Most HR tech content hides behind vague claims and broad benefits. Firstwork is solving for a precise problem, revolving around frontline workers stuck in an activation gap that costs employers candidates and revenue every day.

We say exactly what it does and exactly who it's for. Excited to see what the community thinks.

v cool! nice work team 🙌

Looks awesome, any plans to bring it to other CAD software or is Fusion the primary focus for now?

Beautiful product, the product had really evolved into something admirable.
I have been using the Annotation feature a lot lately, I generally record tutorials and documentation videos. Would there be any support for annotations during the recording as well?

Hey, this looks nice!
Will the transcript-driven editing, text-to-speech and ai-captions work with other languages like german?

Huge congrats on the launch! 🎉 I’ve been using Dina since the very first days and it’s hands-down one of the best screen recorders I’ve tried. The pace of development has been super fast, the responses from Zaid are always on point, and the overall experience feels incredibly polished.
Everything just works.

Excited to keep using it and see where you take it next!

Looks cool! While I appreciate the callout on the resolution export, can you share more about the codec / bitrate that Dina exports? All of these tools fall apart for me in the same way at the end: the export looks terrible because of the default and non-adjustable export settings.

Hi everyone!

Game development is a very context-heavy environment, and that is exactly where many generic coding agents struggle.

@Unity building its own AI layer makes a lot of sense. The assistant sits inside the Editor, understands Unity-specific workflows, and can work with project context like scenes, GameObjects, components, assets, and Editor actions. That is the kind of environment a game dev agent actually needs.

You can use Unity’s own agent, or connect a supported third-party AI subscription through AI Gateway and MCP. The suite also includes control surfaces like undo, AI-generated asset tagging, permissions, and generator controls, which matter a lot when AI starts touching real projects.

For anyone curious about the training data and model policy side, Unity has a detailed Guiding Principles page.

Hey Product Hunt 👋

I’m Manuel Tumiati, co-founder and CTO of Zyphe. Thanks @rajiv_ayyangar for the hunt. Excited to be part of Pitch by Deel NYC.

KYC and KYB are still broken.

People upload the same documents again and again across different platforms, teams store sensitive data they do not want to hold, and compliance teams waste huge amounts of time on repetitive checks.

So we built Zyphe.

Zyphe is the KYC and KYB platform that lets teams and AI agents run identity checks without storing personal data, while making verification reusable across platforms and workflows.

Verify once, reuse everywhere.

That means:

🔐 Less sensitive data stored

♻️ No more repeating the same verification with the same documents across platforms

🤖 Agentic compliance that lets AI agents safely handle parts of compliance workflows

💸 Lower compliance costs and much less manual effort

The core idea is simple:
compliance should be private, reusable, and automated.

Would love your feedback on 3 things:

1. What is the most painful part of KYC or KYB today?

2. Does "verify once, reuse everywhere" resonate?

3. Would you trust AI agents to handle part of compliance?

We’ll be here all day. Ask us anything

Does the KYC Passport require a separate and dedicated app to use it?

Talk about the right timing... feels like there's a new government (state, federal, whatever) talking about age gating and identity verification every single day. Scary to think that a lot of these sites are responsible for securing that information (or even if they trust a third party, those third parties storing data in a centralized way). Good stuff!

Where does my passport photo and selfie actually live after I upload it? "Decentralized" sounds cool but I want to know who can see it.

What does it mean "improve their mental wellbeing"?

Is this focused primarily on healthy folks or those with chronic illness? I could see their being a significant market for those with brain fog and the like if it has actually been tested in that population.

Hey PH!

We're building a world model for oncology. AI that can predict and simulate disease progression based on a patient's full history across imaging, treatment, and clinical data.

That's our vision. Today, we're shipping clinical AI tools for medical imaging that serve as the building blocks (for example, automated body-composition analysis, tumor detection and tumor characterization) toward that vision. These are already powering studies in leading US hospitals, and they're how we're generating revenue while we build toward something bigger.

Medical imaging AI today looks at a single scan in isolation. It doesn't account for prior scans, treatment history, or how the disease has evolved. Radiologists carry that longitudinal understanding in their heads, but their software doesn't. We think that's one of the fundamental limitations holding back AI in cancer care.

I'm Luca and my co-founder is Angelica. We started working together in 2023 at Stanford's Department of Medicine. Between that and my time on Apple's Health AI team, we kept seeing world-class research that never reached the tools clinicians actually use. Nucleo exists to close that gap!

Hi everyone!

Steam Controller is very Valve in the best way.

TMR thumbsticks, trackpads, gyro, Grip Sense, four back buttons, community layouts, per-game configs, and a wireless puck that also works as a magnetic charger.

Every part of it is designed around one idea: play more of your Steam library without forcing every game into the same input shape.

Popped open @Peloton today and got this promo:

So I was like... why haven't I heard of this before? Then decided to check Product Hunt, and whaddya know, @max_gomez launched @Breathwrk here six years ago!

Looks like Peloton paid $2.2M for the app, announcing the acquisition last fall. Appears to finally be rolling out.

the peloton acquisition context is interesting because it suggests breathwork is finally being treated as a first-class recovery and performance tool rather than a wellness afterthought. box breathing and physiological sighs have solid research behind them but most people never build the habit because there's no obvious moment to start. the streak and reminder system is probably doing more work here than the exercises themselves.

Big fan of products that make a single boring-but-high-leverage habit effortless — breathwork sits in the same category as drinking enough water or eating consistently during deal sprints. The Peloton angle is interesting because it suggests a wellness stack play (movement → breath → nutrition) rather than another standalone app fighting for screen time. I worked on a similar problem on the food side with DishRoll, an AI weekly meal planner I built after years of skipping lunch during M&A closings. Curious whether you see Breathwrk eventually integrating with HealthKit / wearables passively (HRV, stress) to trigger sessions, rather than relying on the user remembering to open the app?

The "context layer" framing is exactly what's missing in most production agents — the value isn't "call this API", it's "reason across systems without rebuilding the join at runtime." 80% fewer tokens is a serious unlock when you're deploying agents on noisy, time-sensitive data streams. We hit a parallel problem on the financial side: agents watching prediction markets and trading flows produce signal soup unless you give them a unified context of positions, baselines, and event history. That's basically what we built into PolyMind for Polymarket alerts. Curious how Airbyte's Context Store handles temporal data — do you snapshot events with timestamps so an agent can reason about "what changed since 9am" without re-pulling raw API history?

I feel like a lot of these products are solving for invented solutions.

I had limited coding experience before starting to build my own enterprise apps + websites in Feb, and I have honestly not come up against I problem I have not been able to solve on my own.

By the time I knew N8n existed, I had already automated all of the ops in one of my projects.

I feel like having something of our the box for connectors could be useful but also as a discovery exercise, maybe part of your service could be to introduce new endpoints that customers previously had not considered (i.e. why not use X instead of Y).

Really like the focus on next actions .A lot of finance tools show data but very few actually guide behavior. That’s where real value is.

Great idea, but how do you define what a user should do next? I.e. Does it also show users what types of income are trending and what areas they should focus more on? Would be helpful for paying down bills if users could track X amount towards a total sum, X payments left - Make more payments to save X amount on interest.

Hey everyone,
I'm Filippo, the founder of Blaze, and I'm genuinely pumped to be launching here today. 🚀

I built Blaze because I couldn't find a single app that did everything I needed without making me feel like I was managing the app instead of my actual work. So I built one.

What started as a personal tool is now something I use every single day and I'm so excited to finally put it in your hands.

A few things I'd love your honest take on:
- What's the first thing you do when you open your calendar every morning?
- What feature would make you actually switch from your current setup?

I read every single comment and it goes straight into the roadmap. This is day one for us here, and your feedback shapes what comes next.

The first 1,000 subscribers can enjoy Blaze Pro free of charge for 2 months. You can redeem the offer via this link 👉 https://apps.apple.com/redeem?ctx=offercodes&id=6756891022&code=PRODUCTHUNT

Thanks for being here. Let's make productivity not suck. 💪

Is there a limit on how many calendars it can connect?
I always find it very hard to manage all my work + family calendars.

The hardest part of AI scheduling is it doesn't know what actually takes energy vs what's just clock time. Does it learn from how you actually spend your day or mostly from calendar data? As a solo dev, context switching is the real killer.